WinDOM: Self-Family Distillation for Small-Model GUI Grounding

Small ($\sim$2B) GUI-grounding agents are attractive for on-device deployment, accessibility tooling, and low-cost iteration, but at this scale they face two open recipe questions: how to obtain bounding-box training data without expensive human annotation, and how to combine supervised fine-tuning with reinforcement learning. We address both, with the explicit goal of pushing small-model performance rather than scaling up. WinDOM is a $54{,}425$-record grounding corpus harvested by driving an open-source Windows 11 web reimplementation under headless Playwright, with bounding boxes read directly off the DOM and no OCR or human annotation. Self-Family Distillation (SFD) is a single rejection-sampling cold-start parameterised only by the teacher choice: either an EMA of the student (no external model) or a frozen larger same-family teacher. We then treat the saturation depth of the SFD cold-start as an explicit GRPO hyperparameter. On a Qwen3.5-2B student, the under-saturated cold-start is a better GRPO initialiser than the converged one: SFD-4B with Early-init RL gains $+5.4$ OOD-mean ($+3.5$ ScreenSpot-Pro, $+7.0$ OSWorld-G, $+5.8$ ScreenSpot-V2) over the base. The same-size EMA mode lands within roughly one OOD-mean point of the cross-size $4$B variant ($65.2$ vs $66.3$) without an external teacher.

Robot Navigation in a Crowd by Integrating Deep Reinforcement Learning and Online Planning

It is still an open and challenging problem for mobile robots navigating along time-efficient and collision-free paths in a crowd. The main challenge comes from the complex and sophisticated interaction mechanism, which requires the robot to understand the crowd and perform proactive and foresighted behaviors. Deep reinforcement learning is a promising solution to this problem. However, most previous learning methods incur a tremendous computational burden. To address these problems, we propose a graph-based deep reinforcement learning method, SG-DQN, that (i) introduces a social attention mechanism to extract an efficient graph representation for the crowd-robot state; (ii) directly evaluates the coarse q-values of the raw state with a learned dueling deep Q network(DQN); and then (iii) refines the coarse q-values via online planning on possible future trajectories. The experimental results indicate that our model can help the robot better understand the crowd and achieve a high success rate of more than 0.99 in the crowd navigation task. Compared against previous state-of-the-art algorithms, our algorithm achieves an equivalent, if not better, performance while requiring less than half of the computational cost.